Research On Human Pose Estimation Model And Its Transfer To Cardiac Landmarks Localization

Relying on the booming of "big data" technology and high-performance processors,deep learning has become a dominant research method in the field of artificial intelligence,such as computer vision and natural language processing,with its strong robustness and effectiveness.Human pose estimation is a very challenging research in computer vision.It can be used for: motion recognition,abnormal behavior detection,security and so on.In the field of cardiac medical image processing,cardiac landmarks locating is also a significant study.Accurate detection result helps to achieve automatic ventricle segmentation applications,it is also a necessary condition for many subsequent image processing and analysis steps.There is an inherent connection between the two tasks in terms of form and characteristics.The purpose of this paper is to propose a model based on deep learning to improve human pose estimation.At the same time,this paper analyzes the specific differences and similarities between these two tasks,and uses the method of transfer learning to apply the new model to the task of cardiac landmarks locating.In the study of cardiac landmarks locating,there are relatively few methods based on deep learning.However,in the field of human pose estimation,methods based on deep learning have emerged in an endless stream.Therefore,this paper begins with the task of human pose estimation and analyzes the evolution of the deep learning model in this task.This task currently has problems of scale differences between human joints and redundant information interference.For the former,this paper introduces the Feature pyramid network popular in the field of object detection,and solves this problem by extracting features from multi-scale feature maps and merging high-level semantic features with underlying image features.For the latter,this paper will design an attention model based on a mask and integrate it into the network structure.The attention model is widely used in image caption,image recognition and other fields.Its purpose is to simulate the mechanism of the human visual system to pay more attention to the main task and filter the redundant information.This article adds a comparison experiment after each improvement to verify the hypothesis.Finally,a new model—the feature pyramid attention hourglass network—was proposed,which achieved the best detection effect on MPII and LSP human pose estimation datasets.The last part of this paper analyzes the differences and connections between the cardiac landmarks locating and human pose estimation.These differences and connections correspond to the problems and basic principles that transfer learning can solve.Subsequently,the model-based migration learning method was used to apply the new model to the task of locating the cardiac landmarks,and a comparative experiment was performed to verify the effect of the model.Experiments show that the new model has achieved the best results in the cardiac landmark locating task.The Pyramid attention network model has achieved good results in human pose estimation and cardiac landmarks locating,which contributes to the development of related fields and advances in medical technology.